On trials in which a neuron tuned for upward
motion fired more than its average, the monkey was more likely to report seeing upward than downward motion. Since that initial study, correlations between the fluctuations in the responses of individual neurons and behavior (typically called choice probability for discrimination tasks or detect probability for detection tasks) have been observed in a variety of sensory areas and behavioral tasks (for review, see Nienborg et al., 2012 and Parker and Newsome, selleck inhibitor 1998). The existence of such neuron-behavior correlations, when combined with data from more causal experimental methods like pharmacology, lesions, or electrical stimulation, can provide evidence that those neurons are part of the neural mechanisms underlying specific percepts or behaviors (Parker and Newsome, 1998). Using neuron-behavior correlations (or other experimental methods) to infer the computation that downstream areas perform to decode sensory information from areas like
MT has been much more difficult, however. Selleckchem Akt inhibitor This difficulty has at least three sources. (1) The relationship between any one neuron’s activity and behavior is typically weak and noisy. This is expected because a large number of neurons in multiple brain areas likely contribute to any behavior, but it makes neuron-behavior correlations difficult to measure and interpret. (2) Neuron-behavior correlations are highly influenced by, and in some cases arise solely because of, variability that is shared among groups of neurons (Nienborg and Cumming, 2010). If the firing rates of many neurons rise and fall together, the responses of any one neuron will
be correlated with behavior because its fluctuations reflect the activity of a large population. (Such shared variability is typically quantified as correlations between the trial-by-trial fluctuations between pairs of neurons and referred to as spike count correlation or noise correlation.) This shared variability makes it possible to observe neuron-behavior correlations, but it can also make such correlations arise artifactually: a neuron’s response may be correlated with behavior even if it is not involved in the why underlying computation if its variability is shared with neurons that contribute to the behavior. (3) Neuron-behavior correlations are influenced by variability in external factors such as the visual stimuli used, the difficulty of the task, or aspects of the animal’s cognitive state such as its motivation level. Because neuron-behavior correlations are typically measured in one neuron per experimental session, day-to-day variability in these factors might cloud the dependence of these measurements on factors such as the neuron’s tuning. These problems can be mitigated by using an experimental system for which the stimuli, psychophysical task, sensory responses, motor system, and behavioral output have been well characterized.